Independent work. Smarter tools. Better business.The Freelance Guruji journal
Technology

Can ChatGPT Transcribe Audio? Check the Workflow Before Uploading

Editorial illustration of a microphone and audio signals representing Can ChatGPT Transcribe Audio? Check the Workflow Before Uploading.

A client sends you an audio file — an interview, a meeting, a voice note — and asks for a transcript. You have heard that ChatGPT can do almost anything with text, so it seems reasonable to ask whether it can transcribe audio too. The short answer is that it is not that simple, and the longer answer matters, because uploading audio to the wrong place can waste your time, cost you money, leak private information, or produce a transcript you cannot trust.

Transcription is not one job. It is a chain of jobs: getting the audio into a usable form, turning speech into text, checking the text against the audio, and then doing whatever the transcript is for — a summary, a set of quotes, an article, a set of notes. ChatGPT is good at some parts of that chain and not involved in others at all. Before you upload anything, work out which part you need help with, and check whether the tool you are about to use actually does that part.

Start with the question the client actually asked

When someone asks “can you transcribe this?”, they rarely mean “turn every word into text and send it back”. They usually mean something more specific, and the specific thing changes the work completely. Do they need a word-for-word transcript, where every pause and false start matters? Do they need a clean version, with the ums and repetitions removed? Do they need a summary, with the main points and the decisions? Do they need quotes they can publish, with the speaker identified? Each of those is a different job, with a different level of accuracy needed, and a different amount of checking required.

So ask before you upload. “Do you need every word, or a clean summary? Who will read it, and what will they use it for? How accurate does it need to be — is this a legal record, or a set of notes for yourself?” A transcript for a legal record needs a different standard of checking from a transcript for your own notes, and a transcript that will be published needs the speakers to agree to it. Do not assume. The answer to those questions decides the whole workflow, including whether an automated tool is suitable at all.

Understand what ChatGPT does and does not do

ChatGPT is a text tool. You give it text, and it gives you text back. It does not listen to audio in the way a transcription service does. Some versions and apps built around it can accept an audio file and produce text, but that is not the same as ChatGPT itself transcribing, and the difference matters for your workflow. When an audio file goes into a tool, something has to turn the sound into words first — that step is called speech-to-text, or automatic speech recognition — and only after that step does a language tool have any text to work with.

That means there are really two questions to ask, not one. First, what turns the audio into text? Second, what happens to the text after that? The first question is about accuracy: how well does the tool hear different voices, accents, background noise and technical words? The second question is about what you do with the text: clean it up, summarise it, pull out quotes, translate it. ChatGPT is useful for the second question. For the first, you need a transcription step, and you need to know which one you are using and how good it is.

Check the audio before you check the tool

A transcription tool cannot fix bad audio. If the recording is quiet, muffled, full of background noise, or has two people talking over each other, the transcript will be wrong in exactly the places that matter — names, numbers, decisions, dates. No amount of clever prompting will recover words that were never audible in the first place.

Before you upload, listen to a few minutes of the audio yourself. Ask: can I hear every speaker clearly? Is there background noise — traffic, music, a fan? Do people talk over each other? Are there long sections where nobody is speaking? Is the file in a format the tool accepts, and is it short enough for the tool’s limits? If the answer to any of those is a problem, fix the audio first if you can, or tell the client that the transcript will have gaps and that those gaps need to be marked, not guessed at. A transcript that invents words for the parts it could not hear is worse than a transcript with gaps, because the gaps are honest and the invented words look true.

Privacy: do not upload what you should not share

This is the step people skip, and it is the one with the most serious consequences. An audio file can contain private information: a client’s business plans, a person’s health details, a conversation that was never meant to be shared, the names of people who did not agree to be recorded. Uploading that file to an online service means sending it to someone else’s servers, under someone else’s terms, and you need to know what those terms say before you send it.

Before you upload anything, ask three questions. First, who owns this audio, and do I have permission to share it with a third-party service? Second, what does the service do with the file — does it store it, use it to train its models, share it with anyone else, and can I delete it afterwards? Third, is there a safer route — a tool that runs on your own computer, or a service with a written agreement about how it handles data? If you cannot answer those questions, do not upload the file. Tell the client why, and offer the safer route. A client who understands that you protected their recording will trust you more, not less.

Accuracy: check the transcript against the audio

Automated transcription is good, and it is not perfect. It makes mistakes with names, with numbers, with words that sound alike, and with anything said quickly or quietly. It also struggles with multiple speakers, and it can assign the wrong words to the wrong person. That matters a great deal if the transcript is going to be quoted, published, or used to make a decision.

So never send a transcript you have not checked. Listen to the audio while you read the transcript, at least for the important parts — the names, the numbers, the decisions, the quotes you plan to use. Mark anything you are unsure about, and do not guess at it. If a word is unclear, write it as unclear, or go back and listen again. If two speakers sound alike, check who said what. A transcript with a few marked gaps is a useful document. A transcript that looks complete but is wrong in the important places is a trap, and it will be quoted as if it were true.

The workflow that actually works

Here is a workflow that works for most freelance jobs, using the tools in the order they are needed. First, agree with the client what the transcript is for and how accurate it needs to be. Second, check the audio — listen to a sample, check the format and the length, and note any problems. Third, choose a transcription step that fits the privacy rules for this file, and run the audio through it. Fourth, check the transcript against the audio, fixing names, numbers and speaker labels, and marking anything unclear. Fifth, only then use a language tool to clean up, summarise or reformat the text, and check that output too, because a summary can change the meaning of what was said.

That order matters. If you skip the checking step and go straight from audio to summary, you will produce a confident summary of a transcript that may be wrong, and nobody will know until someone checks the original recording. The checking step is the work. The tools are just the tools.

What to tell the client before you start

A short message before you start prevents most of the problems later. Tell the client what you will deliver, how long it will take, and what the limits are. “I will deliver a clean transcript with speaker labels, checked against the audio for names and numbers. Sections where the audio was unclear will be marked as unclear, not guessed.”

A worked example: a one-hour interview, done properly (illustration)

This is a made-up example with simple round numbers, purely to show the shape of the work — not a real client and not a recommended price.

A freelancer is asked to transcribe a one-hour interview with two speakers, recorded on a phone in a quiet room. The client wants a clean transcript with speaker labels, to use for an article. The freelancer listens to the first five minutes, confirms both speakers are clear, and checks that the file is in a format the transcription step accepts. The transcription step takes about an hour of processing for the hour of audio, and produces a draft transcript. The freelancer then listens to the audio while reading the draft, which takes about ninety minutes, fixing names, correcting numbers, and marking two sections where a speaker mumbled as unclear. Only then does the freelancer use a language tool to tidy the grammar and remove repeated words, and checks that tidy version against the checked transcript. The total work is about three hours for one hour of audio — and most of that time is checking, not transcribing. A freelancer who promises a one-hour transcript for one hour of audio is promising a draft, not a finished transcript, and the client should know the difference.

Quick checklist before you upload any audio

  • I know what the transcript is for, and how accurate it needs to be.
  • I have listened to a sample of the audio and noted any problems.
  • I have permission to share this audio with a third-party service, and I know what the service does with it.
  • I know which step turns the audio into text, and which step works with the text afterwards.
  • I will check the transcript against the audio, especially names, numbers and quotes.
  • Unclear words will be marked as unclear, not guessed.
  • I have told the client what I will deliver, what the limits are, and how their file will be handled.

If you can tick all seven, you are ready to upload. If you cannot, the missing item is the work to do first — and it is better to do it now than to explain a leaked recording or a wrong quote later.